13 #include "lib/axmap.h"
18 struct io_completion_data {
21 int error; /* output */
22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
23 struct timeval time; /* output */
27 * The ->io_axmap contains a map of blocks we have or have not done io
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
30 static bool random_map_free(struct fio_file *f, const uint64_t block)
32 return !axmap_isset(f->io_axmap, block);
36 * Mark a given offset as used in the map.
38 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40 unsigned int min_bs = td->o.rw_min_bs;
41 struct fio_file *f = io_u->file;
42 unsigned int nr_blocks;
45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
48 if (!(io_u->flags & IO_U_F_BUSY_OK))
49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
55 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
61 assert(ddir_rw(ddir));
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
66 max_size = f->io_size;
67 if (max_size > f->real_file_size)
68 max_size = f->real_file_size;
71 max_size = td->o.zone_range;
73 if (td->o.min_bs[ddir] > td->o.ba[ddir])
74 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
76 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
84 struct flist_head list;
88 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
89 enum fio_ddir ddir, uint64_t *b)
93 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
94 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
95 uint64_t frand_max, lastb;
97 lastb = last_block(td, f, ddir);
101 frand_max = rand_max(&td->random_state);
102 r = __rand(&td->random_state);
104 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
106 *b = lastb * (r / ((uint64_t) frand_max + 1.0));
110 assert(fio_file_lfsr(f));
112 if (lfsr_next(&f->lfsr, &off))
119 * if we are not maintaining a random map, we are done.
121 if (!file_randommap(td, f))
125 * calculate map offset and check if it's free
127 if (random_map_free(f, *b))
130 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
131 (unsigned long long) *b);
133 *b = axmap_next_free(f->io_axmap, *b);
134 if (*b == (uint64_t) -1ULL)
140 static int __get_next_rand_offset_zipf(struct thread_data *td,
141 struct fio_file *f, enum fio_ddir ddir,
144 *b = zipf_next(&f->zipf);
148 static int __get_next_rand_offset_pareto(struct thread_data *td,
149 struct fio_file *f, enum fio_ddir ddir,
152 *b = pareto_next(&f->zipf);
156 static int __get_next_rand_offset_gauss(struct thread_data *td,
157 struct fio_file *f, enum fio_ddir ddir,
160 *b = gauss_next(&f->gauss);
165 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
167 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
168 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
170 return r1->off - r2->off;
173 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
174 enum fio_ddir ddir, uint64_t *b)
176 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
177 return __get_next_rand_offset(td, f, ddir, b);
178 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
179 return __get_next_rand_offset_zipf(td, f, ddir, b);
180 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
181 return __get_next_rand_offset_pareto(td, f, ddir, b);
182 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
183 return __get_next_rand_offset_gauss(td, f, ddir, b);
185 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
190 * Sort the reads for a verify phase in batches of verifysort_nr, if
193 static inline bool should_sort_io(struct thread_data *td)
195 if (!td->o.verifysort_nr || !td->o.do_verify)
199 if (td->runstate != TD_VERIFYING)
201 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
202 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
208 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
214 if (td->o.perc_rand[ddir] == 100)
217 frand_max = rand_max(&td->seq_rand_state[ddir]);
218 r = __rand(&td->seq_rand_state[ddir]);
219 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
221 return v <= td->o.perc_rand[ddir];
224 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
225 enum fio_ddir ddir, uint64_t *b)
230 if (!should_sort_io(td))
231 return get_off_from_method(td, f, ddir, b);
233 if (!flist_empty(&td->next_rand_list)) {
235 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
242 for (i = 0; i < td->o.verifysort_nr; i++) {
243 r = malloc(sizeof(*r));
245 ret = get_off_from_method(td, f, ddir, &r->off);
251 flist_add(&r->list, &td->next_rand_list);
257 assert(!flist_empty(&td->next_rand_list));
258 flist_sort(NULL, &td->next_rand_list, flist_cmp);
262 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
263 enum fio_ddir ddir, uint64_t *b)
265 if (!get_next_rand_offset(td, f, ddir, b))
268 if (td->o.time_based) {
269 fio_file_reset(td, f);
270 if (!get_next_rand_offset(td, f, ddir, b))
274 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
275 f->file_name, (unsigned long long) f->last_pos[ddir],
276 (unsigned long long) f->real_file_size);
280 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
281 enum fio_ddir ddir, uint64_t *offset)
283 struct thread_options *o = &td->o;
285 assert(ddir_rw(ddir));
287 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
289 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
291 if (f->last_pos[ddir] < f->real_file_size) {
294 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
295 f->last_pos[ddir] = f->real_file_size;
297 pos = f->last_pos[ddir] - f->file_offset;
298 if (pos && o->ddir_seq_add) {
299 pos += o->ddir_seq_add;
302 * If we reach beyond the end of the file
303 * with holed IO, wrap around to the
306 if (pos >= f->real_file_size)
307 pos = f->file_offset;
317 static int get_next_block(struct thread_data *td, struct io_u *io_u,
318 enum fio_ddir ddir, int rw_seq,
319 unsigned int *is_random)
321 struct fio_file *f = io_u->file;
325 assert(ddir_rw(ddir));
331 if (should_do_random(td, ddir)) {
332 ret = get_next_rand_block(td, f, ddir, &b);
336 io_u_set(io_u, IO_U_F_BUSY_OK);
337 ret = get_next_seq_offset(td, f, ddir, &offset);
339 ret = get_next_rand_block(td, f, ddir, &b);
343 ret = get_next_seq_offset(td, f, ddir, &offset);
346 io_u_set(io_u, IO_U_F_BUSY_OK);
349 if (td->o.rw_seq == RW_SEQ_SEQ) {
350 ret = get_next_seq_offset(td, f, ddir, &offset);
352 ret = get_next_rand_block(td, f, ddir, &b);
355 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
356 if (f->last_start[ddir] != -1ULL)
357 offset = f->last_start[ddir] - f->file_offset;
362 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
369 io_u->offset = offset;
371 io_u->offset = b * td->o.ba[ddir];
373 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
382 * For random io, generate a random new block and see if it's used. Repeat
383 * until we find a free one. For sequential io, just return the end of
384 * the last io issued.
386 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
387 unsigned int *is_random)
389 struct fio_file *f = io_u->file;
390 enum fio_ddir ddir = io_u->ddir;
393 assert(ddir_rw(ddir));
395 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
397 td->ddir_seq_nr = td->o.ddir_seq_nr;
400 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
403 if (io_u->offset >= f->io_size) {
404 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
405 (unsigned long long) io_u->offset,
406 (unsigned long long) f->io_size);
410 io_u->offset += f->file_offset;
411 if (io_u->offset >= f->real_file_size) {
412 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
413 (unsigned long long) io_u->offset,
414 (unsigned long long) f->real_file_size);
421 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
422 unsigned int *is_random)
424 if (td->flags & TD_F_PROFILE_OPS) {
425 struct prof_io_ops *ops = &td->prof_io_ops;
427 if (ops->fill_io_u_off)
428 return ops->fill_io_u_off(td, io_u, is_random);
431 return __get_next_offset(td, io_u, is_random);
434 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
437 struct fio_file *f = io_u->file;
439 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
442 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
443 unsigned int is_random)
445 int ddir = io_u->ddir;
446 unsigned int buflen = 0;
447 unsigned int minbs, maxbs;
451 assert(ddir_rw(ddir));
453 if (td->o.bs_is_seq_rand)
454 ddir = is_random ? DDIR_WRITE: DDIR_READ;
456 minbs = td->o.min_bs[ddir];
457 maxbs = td->o.max_bs[ddir];
463 * If we can't satisfy the min block size from here, then fail
465 if (!io_u_fits(td, io_u, minbs))
468 frand_max = rand_max(&td->bsrange_state);
470 r = __rand(&td->bsrange_state);
472 if (!td->o.bssplit_nr[ddir]) {
473 buflen = 1 + (unsigned int) ((double) maxbs *
474 (r / (frand_max + 1.0)));
481 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
482 struct bssplit *bsp = &td->o.bssplit[ddir][i];
486 if ((r <= ((frand_max / 100L) * perc)) &&
487 io_u_fits(td, io_u, buflen))
492 if (td->o.verify != VERIFY_NONE)
493 buflen = (buflen + td->o.verify_interval - 1) &
494 ~(td->o.verify_interval - 1);
496 if (!td->o.bs_unaligned && is_power_of_2(minbs))
497 buflen &= ~(minbs - 1);
499 } while (!io_u_fits(td, io_u, buflen));
504 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
505 unsigned int is_random)
507 if (td->flags & TD_F_PROFILE_OPS) {
508 struct prof_io_ops *ops = &td->prof_io_ops;
510 if (ops->fill_io_u_size)
511 return ops->fill_io_u_size(td, io_u, is_random);
514 return __get_next_buflen(td, io_u, is_random);
517 static void set_rwmix_bytes(struct thread_data *td)
522 * we do time or byte based switch. this is needed because
523 * buffered writes may issue a lot quicker than they complete,
524 * whereas reads do not.
526 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
527 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
530 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
532 uint64_t frand_max = rand_max(&td->rwmix_state);
536 r = __rand(&td->rwmix_state);
537 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
539 if (v <= td->o.rwmix[DDIR_READ])
545 int io_u_quiesce(struct thread_data *td)
550 * We are going to sleep, ensure that we flush anything pending as
551 * not to skew our latency numbers.
553 * Changed to only monitor 'in flight' requests here instead of the
554 * td->cur_depth, b/c td->cur_depth does not accurately represent
555 * io's that have been actually submitted to an async engine,
556 * and cur_depth is meaningless for sync engines.
558 if (td->io_u_queued || td->cur_depth) {
561 ret = td_io_commit(td);
564 while (td->io_u_in_flight) {
567 ret = io_u_queued_complete(td, 1);
575 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
577 enum fio_ddir odir = ddir ^ 1;
580 assert(ddir_rw(ddir));
581 now = utime_since_now(&td->start);
584 * if rate_next_io_time is in the past, need to catch up to rate
586 if (td->rate_next_io_time[ddir] <= now)
590 * We are ahead of rate in this direction. See if we
593 if (td_rw(td) && td->o.rwmix[odir]) {
595 * Other direction is behind rate, switch
597 if (td->rate_next_io_time[odir] <= now)
601 * Both directions are ahead of rate. sleep the min
602 * switch if necissary
604 if (td->rate_next_io_time[ddir] <=
605 td->rate_next_io_time[odir]) {
606 usec = td->rate_next_io_time[ddir] - now;
608 usec = td->rate_next_io_time[odir] - now;
612 usec = td->rate_next_io_time[ddir] - now;
614 if (td->o.io_submit_mode == IO_MODE_INLINE)
617 usec = usec_sleep(td, usec);
623 * Return the data direction for the next io_u. If the job is a
624 * mixed read/write workload, check the rwmix cycle and switch if
627 static enum fio_ddir get_rw_ddir(struct thread_data *td)
632 * see if it's time to fsync
634 if (td->o.fsync_blocks &&
635 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
636 td->io_issues[DDIR_WRITE] && should_fsync(td))
640 * see if it's time to fdatasync
642 if (td->o.fdatasync_blocks &&
643 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
644 td->io_issues[DDIR_WRITE] && should_fsync(td))
645 return DDIR_DATASYNC;
648 * see if it's time to sync_file_range
650 if (td->sync_file_range_nr &&
651 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
652 td->io_issues[DDIR_WRITE] && should_fsync(td))
653 return DDIR_SYNC_FILE_RANGE;
657 * Check if it's time to seed a new data direction.
659 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
661 * Put a top limit on how many bytes we do for
662 * one data direction, to avoid overflowing the
665 ddir = get_rand_ddir(td);
667 if (ddir != td->rwmix_ddir)
670 td->rwmix_ddir = ddir;
672 ddir = td->rwmix_ddir;
673 } else if (td_read(td))
675 else if (td_write(td))
680 td->rwmix_ddir = rate_ddir(td, ddir);
681 return td->rwmix_ddir;
684 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
686 enum fio_ddir ddir = get_rw_ddir(td);
688 if (td_trimwrite(td)) {
689 struct fio_file *f = io_u->file;
690 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
696 io_u->ddir = io_u->acct_ddir = ddir;
698 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
699 td->o.barrier_blocks &&
700 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
701 td->io_issues[DDIR_WRITE])
702 io_u_set(io_u, IO_U_F_BARRIER);
705 void put_file_log(struct thread_data *td, struct fio_file *f)
707 unsigned int ret = put_file(td, f);
710 td_verror(td, ret, "file close");
713 void put_io_u(struct thread_data *td, struct io_u *io_u)
720 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
721 put_file_log(td, io_u->file);
724 io_u_set(io_u, IO_U_F_FREE);
726 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
728 assert(!(td->flags & TD_F_CHILD));
730 io_u_qpush(&td->io_u_freelist, io_u);
732 td_io_u_free_notify(td);
735 void clear_io_u(struct thread_data *td, struct io_u *io_u)
737 io_u_clear(io_u, IO_U_F_FLIGHT);
741 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
743 struct io_u *__io_u = *io_u;
744 enum fio_ddir ddir = acct_ddir(__io_u);
746 dprint(FD_IO, "requeue %p\n", __io_u);
753 io_u_set(__io_u, IO_U_F_FREE);
754 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
755 td->io_issues[ddir]--;
757 io_u_clear(__io_u, IO_U_F_FLIGHT);
758 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
760 assert(!(td->flags & TD_F_CHILD));
763 io_u_rpush(&td->io_u_requeues, __io_u);
765 td_io_u_free_notify(td);
769 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
771 unsigned int is_random;
773 if (td->io_ops->flags & FIO_NOIO)
776 set_rw_ddir(td, io_u);
779 * fsync() or fdatasync() or trim etc, we are done
781 if (!ddir_rw(io_u->ddir))
785 * See if it's time to switch to a new zone
787 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
788 struct fio_file *f = io_u->file;
791 f->file_offset += td->o.zone_range + td->o.zone_skip;
794 * Wrap from the beginning, if we exceed the file size
796 if (f->file_offset >= f->real_file_size)
797 f->file_offset = f->real_file_size - f->file_offset;
798 f->last_pos[io_u->ddir] = f->file_offset;
799 td->io_skip_bytes += td->o.zone_skip;
803 * No log, let the seq/rand engine retrieve the next buflen and
806 if (get_next_offset(td, io_u, &is_random)) {
807 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
811 io_u->buflen = get_next_buflen(td, io_u, is_random);
813 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
817 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
818 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
819 dprint(FD_IO, " off=%llu/%lu > %llu\n",
820 (unsigned long long) io_u->offset, io_u->buflen,
821 (unsigned long long) io_u->file->real_file_size);
826 * mark entry before potentially trimming io_u
828 if (td_random(td) && file_randommap(td, io_u->file))
829 mark_random_map(td, io_u);
832 dprint_io_u(io_u, "fill_io_u");
833 td->zone_bytes += io_u->buflen;
837 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
866 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
868 __io_u_mark_map(td->ts.io_u_submit, nr);
869 td->ts.total_submit++;
872 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
874 __io_u_mark_map(td->ts.io_u_complete, nr);
875 td->ts.total_complete++;
878 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
882 switch (td->cur_depth) {
904 td->ts.io_u_map[idx] += nr;
907 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
944 assert(idx < FIO_IO_U_LAT_U_NR);
945 td->ts.io_u_lat_u[idx]++;
948 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
989 assert(idx < FIO_IO_U_LAT_M_NR);
990 td->ts.io_u_lat_m[idx]++;
993 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
996 io_u_mark_lat_usec(td, usec);
998 io_u_mark_lat_msec(td, usec / 1000);
1002 * Get next file to service by choosing one at random
1004 static struct fio_file *get_next_file_rand(struct thread_data *td,
1005 enum fio_file_flags goodf,
1006 enum fio_file_flags badf)
1008 uint64_t frand_max = rand_max(&td->next_file_state);
1016 r = __rand(&td->next_file_state);
1017 fno = (unsigned int) ((double) td->o.nr_files
1018 * (r / (frand_max + 1.0)));
1021 if (fio_file_done(f))
1024 if (!fio_file_open(f)) {
1027 if (td->nr_open_files >= td->o.open_files)
1028 return ERR_PTR(-EBUSY);
1030 err = td_io_open_file(td, f);
1036 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1037 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1041 td_io_close_file(td, f);
1046 * Get next file to service by doing round robin between all available ones
1048 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1051 unsigned int old_next_file = td->next_file;
1057 f = td->files[td->next_file];
1060 if (td->next_file >= td->o.nr_files)
1063 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1064 if (fio_file_done(f)) {
1069 if (!fio_file_open(f)) {
1072 if (td->nr_open_files >= td->o.open_files)
1073 return ERR_PTR(-EBUSY);
1075 err = td_io_open_file(td, f);
1077 dprint(FD_FILE, "error %d on open of %s\n",
1085 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1087 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1091 td_io_close_file(td, f);
1094 } while (td->next_file != old_next_file);
1096 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1100 static struct fio_file *__get_next_file(struct thread_data *td)
1104 assert(td->o.nr_files <= td->files_index);
1106 if (td->nr_done_files >= td->o.nr_files) {
1107 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1108 " nr_files=%d\n", td->nr_open_files,
1114 f = td->file_service_file;
1115 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1116 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1118 if (td->file_service_left--)
1122 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1123 td->o.file_service_type == FIO_FSERVICE_SEQ)
1124 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1126 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1131 td->file_service_file = f;
1132 td->file_service_left = td->file_service_nr - 1;
1135 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1137 dprint(FD_FILE, "get_next_file: NULL\n");
1141 static struct fio_file *get_next_file(struct thread_data *td)
1143 if (td->flags & TD_F_PROFILE_OPS) {
1144 struct prof_io_ops *ops = &td->prof_io_ops;
1146 if (ops->get_next_file)
1147 return ops->get_next_file(td);
1150 return __get_next_file(td);
1153 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1158 f = get_next_file(td);
1159 if (IS_ERR_OR_NULL(f))
1165 if (!fill_io_u(td, io_u))
1168 put_file_log(td, f);
1169 td_io_close_file(td, f);
1171 fio_file_set_done(f);
1172 td->nr_done_files++;
1173 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1174 td->nr_done_files, td->o.nr_files);
1180 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1181 unsigned long tusec, unsigned long max_usec)
1184 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1185 td_verror(td, ETIMEDOUT, "max latency exceeded");
1186 icd->error = ETIMEDOUT;
1189 static void lat_new_cycle(struct thread_data *td)
1191 fio_gettime(&td->latency_ts, NULL);
1192 td->latency_ios = ddir_rw_sum(td->io_blocks);
1193 td->latency_failed = 0;
1197 * We had an IO outside the latency target. Reduce the queue depth. If we
1198 * are at QD=1, then it's time to give up.
1200 static bool __lat_target_failed(struct thread_data *td)
1202 if (td->latency_qd == 1)
1205 td->latency_qd_high = td->latency_qd;
1207 if (td->latency_qd == td->latency_qd_low)
1208 td->latency_qd_low--;
1210 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1212 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1215 * When we ramp QD down, quiesce existing IO to prevent
1216 * a storm of ramp downs due to pending higher depth.
1223 static bool lat_target_failed(struct thread_data *td)
1225 if (td->o.latency_percentile.u.f == 100.0)
1226 return __lat_target_failed(td);
1228 td->latency_failed++;
1232 void lat_target_init(struct thread_data *td)
1234 td->latency_end_run = 0;
1236 if (td->o.latency_target) {
1237 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1238 fio_gettime(&td->latency_ts, NULL);
1240 td->latency_qd_high = td->o.iodepth;
1241 td->latency_qd_low = 1;
1242 td->latency_ios = ddir_rw_sum(td->io_blocks);
1244 td->latency_qd = td->o.iodepth;
1247 void lat_target_reset(struct thread_data *td)
1249 if (!td->latency_end_run)
1250 lat_target_init(td);
1253 static void lat_target_success(struct thread_data *td)
1255 const unsigned int qd = td->latency_qd;
1256 struct thread_options *o = &td->o;
1258 td->latency_qd_low = td->latency_qd;
1261 * If we haven't failed yet, we double up to a failing value instead
1262 * of bisecting from highest possible queue depth. If we have set
1263 * a limit other than td->o.iodepth, bisect between that.
1265 if (td->latency_qd_high != o->iodepth)
1266 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1268 td->latency_qd *= 2;
1270 if (td->latency_qd > o->iodepth)
1271 td->latency_qd = o->iodepth;
1273 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1276 * Same as last one, we are done. Let it run a latency cycle, so
1277 * we get only the results from the targeted depth.
1279 if (td->latency_qd == qd) {
1280 if (td->latency_end_run) {
1281 dprint(FD_RATE, "We are done\n");
1284 dprint(FD_RATE, "Quiesce and final run\n");
1286 td->latency_end_run = 1;
1287 reset_all_stats(td);
1296 * Check if we can bump the queue depth
1298 void lat_target_check(struct thread_data *td)
1300 uint64_t usec_window;
1304 usec_window = utime_since_now(&td->latency_ts);
1305 if (usec_window < td->o.latency_window)
1308 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1309 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1310 success_ios *= 100.0;
1312 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1314 if (success_ios >= td->o.latency_percentile.u.f)
1315 lat_target_success(td);
1317 __lat_target_failed(td);
1321 * If latency target is enabled, we might be ramping up or down and not
1322 * using the full queue depth available.
1324 bool queue_full(const struct thread_data *td)
1326 const int qempty = io_u_qempty(&td->io_u_freelist);
1330 if (!td->o.latency_target)
1333 return td->cur_depth >= td->latency_qd;
1336 struct io_u *__get_io_u(struct thread_data *td)
1338 struct io_u *io_u = NULL;
1346 if (!io_u_rempty(&td->io_u_requeues))
1347 io_u = io_u_rpop(&td->io_u_requeues);
1348 else if (!queue_full(td)) {
1349 io_u = io_u_qpop(&td->io_u_freelist);
1354 io_u->end_io = NULL;
1358 assert(io_u->flags & IO_U_F_FREE);
1359 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1360 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1364 io_u->acct_ddir = -1;
1366 assert(!(td->flags & TD_F_CHILD));
1367 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1369 } else if (td_async_processing(td)) {
1371 * We ran out, wait for async verify threads to finish and
1374 assert(!(td->flags & TD_F_CHILD));
1375 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1383 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1385 if (!(td->flags & TD_F_TRIM_BACKLOG))
1388 if (td->trim_entries) {
1391 if (td->trim_batch) {
1394 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1395 td->last_ddir != DDIR_READ) {
1396 td->trim_batch = td->o.trim_batch;
1397 if (!td->trim_batch)
1398 td->trim_batch = td->o.trim_backlog;
1402 if (get_trim && !get_next_trim(td, io_u))
1409 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1411 if (!(td->flags & TD_F_VER_BACKLOG))
1414 if (td->io_hist_len) {
1417 if (td->verify_batch)
1419 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1420 td->last_ddir != DDIR_READ) {
1421 td->verify_batch = td->o.verify_batch;
1422 if (!td->verify_batch)
1423 td->verify_batch = td->o.verify_backlog;
1427 if (get_verify && !get_next_verify(td, io_u)) {
1437 * Fill offset and start time into the buffer content, to prevent too
1438 * easy compressible data for simple de-dupe attempts. Do this for every
1439 * 512b block in the range, since that should be the smallest block size
1440 * we can expect from a device.
1442 static void small_content_scramble(struct io_u *io_u)
1444 unsigned int i, nr_blocks = io_u->buflen / 512;
1446 unsigned int offset;
1453 boffset = io_u->offset;
1454 io_u->buf_filled_len = 0;
1456 for (i = 0; i < nr_blocks; i++) {
1458 * Fill the byte offset into a "random" start offset of
1459 * the buffer, given by the product of the usec time
1460 * and the actual offset.
1462 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1463 offset &= ~(sizeof(uint64_t) - 1);
1464 if (offset >= 512 - sizeof(uint64_t))
1465 offset -= sizeof(uint64_t);
1466 memcpy(p + offset, &boffset, sizeof(boffset));
1468 end = p + 512 - sizeof(io_u->start_time);
1469 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1476 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1477 * etc. The returned io_u is fully ready to be prepped and submitted.
1479 struct io_u *get_io_u(struct thread_data *td)
1483 int do_scramble = 0;
1486 io_u = __get_io_u(td);
1488 dprint(FD_IO, "__get_io_u failed\n");
1492 if (check_get_verify(td, io_u))
1494 if (check_get_trim(td, io_u))
1498 * from a requeue, io_u already setup
1504 * If using an iolog, grab next piece if any available.
1506 if (td->flags & TD_F_READ_IOLOG) {
1507 if (read_iolog_get(td, io_u))
1509 } else if (set_io_u_file(td, io_u)) {
1511 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1517 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1521 assert(fio_file_open(f));
1523 if (ddir_rw(io_u->ddir)) {
1524 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1525 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1529 f->last_start[io_u->ddir] = io_u->offset;
1530 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1532 if (io_u->ddir == DDIR_WRITE) {
1533 if (td->flags & TD_F_REFILL_BUFFERS) {
1534 io_u_fill_buffer(td, io_u,
1535 td->o.min_bs[DDIR_WRITE],
1537 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1538 !(td->flags & TD_F_COMPRESS))
1540 if (td->flags & TD_F_VER_NONE) {
1541 populate_verify_io_u(td, io_u);
1544 } else if (io_u->ddir == DDIR_READ) {
1546 * Reset the buf_filled parameters so next time if the
1547 * buffer is used for writes it is refilled.
1549 io_u->buf_filled_len = 0;
1554 * Set io data pointers.
1556 io_u->xfer_buf = io_u->buf;
1557 io_u->xfer_buflen = io_u->buflen;
1561 if (!td_io_prep(td, io_u)) {
1562 if (!td->o.disable_lat)
1563 fio_gettime(&io_u->start_time, NULL);
1565 small_content_scramble(io_u);
1569 dprint(FD_IO, "get_io_u failed\n");
1571 return ERR_PTR(ret);
1574 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1576 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1578 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1581 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1582 io_u->file ? " on file " : "",
1583 io_u->file ? io_u->file->file_name : "",
1584 strerror(io_u->error),
1585 io_ddir_name(io_u->ddir),
1586 io_u->offset, io_u->xfer_buflen);
1588 if (td->io_ops->errdetails) {
1589 char *err = td->io_ops->errdetails(io_u);
1591 log_err("fio: %s\n", err);
1596 td_verror(td, io_u->error, "io_u error");
1599 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1601 __io_u_log_error(td, io_u);
1603 __io_u_log_error(td, io_u);
1606 static inline bool gtod_reduce(struct thread_data *td)
1608 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1609 || td->o.gtod_reduce;
1612 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1613 struct io_completion_data *icd,
1614 const enum fio_ddir idx, unsigned int bytes)
1616 const int no_reduce = !gtod_reduce(td);
1617 unsigned long lusec = 0;
1623 lusec = utime_since(&io_u->issue_time, &icd->time);
1625 if (!td->o.disable_lat) {
1626 unsigned long tusec;
1628 tusec = utime_since(&io_u->start_time, &icd->time);
1629 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1631 if (td->flags & TD_F_PROFILE_OPS) {
1632 struct prof_io_ops *ops = &td->prof_io_ops;
1635 icd->error = ops->io_u_lat(td, tusec);
1638 if (td->o.max_latency && tusec > td->o.max_latency)
1639 lat_fatal(td, icd, tusec, td->o.max_latency);
1640 if (td->o.latency_target && tusec > td->o.latency_target) {
1641 if (lat_target_failed(td))
1642 lat_fatal(td, icd, tusec, td->o.latency_target);
1646 if (!td->o.disable_clat) {
1647 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1648 io_u_mark_latency(td, lusec);
1651 if (!td->o.disable_bw)
1652 add_bw_sample(td, idx, bytes, &icd->time);
1655 add_iops_sample(td, idx, bytes, &icd->time);
1657 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1658 uint32_t *info = io_u_block_info(td, io_u);
1659 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1660 if (io_u->ddir == DDIR_TRIM) {
1661 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1662 BLOCK_INFO_TRIMS(*info) + 1);
1663 } else if (io_u->ddir == DDIR_WRITE) {
1664 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1671 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1672 struct io_completion_data *icd)
1674 struct io_u *io_u = *io_u_ptr;
1675 enum fio_ddir ddir = io_u->ddir;
1676 struct fio_file *f = io_u->file;
1678 dprint_io_u(io_u, "io complete");
1680 assert(io_u->flags & IO_U_F_FLIGHT);
1681 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1684 * Mark IO ok to verify
1688 * Remove errored entry from the verification list
1691 unlog_io_piece(td, io_u);
1693 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1698 if (ddir_sync(ddir)) {
1699 td->last_was_sync = 1;
1701 f->first_write = -1ULL;
1702 f->last_write = -1ULL;
1707 td->last_was_sync = 0;
1708 td->last_ddir = ddir;
1710 if (!io_u->error && ddir_rw(ddir)) {
1711 unsigned int bytes = io_u->buflen - io_u->resid;
1714 td->io_blocks[ddir]++;
1715 td->this_io_blocks[ddir]++;
1716 td->io_bytes[ddir] += bytes;
1718 if (!(io_u->flags & IO_U_F_VER_LIST))
1719 td->this_io_bytes[ddir] += bytes;
1721 if (ddir == DDIR_WRITE) {
1723 if (f->first_write == -1ULL ||
1724 io_u->offset < f->first_write)
1725 f->first_write = io_u->offset;
1726 if (f->last_write == -1ULL ||
1727 ((io_u->offset + bytes) > f->last_write))
1728 f->last_write = io_u->offset + bytes;
1730 if (td->last_write_comp) {
1731 int idx = td->last_write_idx++;
1733 td->last_write_comp[idx] = io_u->offset;
1734 if (td->last_write_idx == td->o.iodepth)
1735 td->last_write_idx = 0;
1739 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1740 td->runstate == TD_VERIFYING))
1741 account_io_completion(td, io_u, icd, ddir, bytes);
1743 icd->bytes_done[ddir] += bytes;
1746 ret = io_u->end_io(td, io_u_ptr);
1748 if (ret && !icd->error)
1751 } else if (io_u->error) {
1752 icd->error = io_u->error;
1753 io_u_log_error(td, io_u);
1756 enum error_type_bit eb = td_error_type(ddir, icd->error);
1758 if (!td_non_fatal_error(td, eb, icd->error))
1762 * If there is a non_fatal error, then add to the error count
1763 * and clear all the errors.
1765 update_error_count(td, icd->error);
1773 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1778 if (!gtod_reduce(td))
1779 fio_gettime(&icd->time, NULL);
1784 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1785 icd->bytes_done[ddir] = 0;
1788 static void ios_completed(struct thread_data *td,
1789 struct io_completion_data *icd)
1794 for (i = 0; i < icd->nr; i++) {
1795 io_u = td->io_ops->event(td, i);
1797 io_completed(td, &io_u, icd);
1805 * Complete a single io_u for the sync engines.
1807 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1809 struct io_completion_data icd;
1812 init_icd(td, &icd, 1);
1813 io_completed(td, &io_u, &icd);
1819 td_verror(td, icd.error, "io_u_sync_complete");
1823 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1824 td->bytes_done[ddir] += icd.bytes_done[ddir];
1830 * Called to complete min_events number of io for the async engines.
1832 int io_u_queued_complete(struct thread_data *td, int min_evts)
1834 struct io_completion_data icd;
1835 struct timespec *tvp = NULL;
1837 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1839 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1843 else if (min_evts > td->cur_depth)
1844 min_evts = td->cur_depth;
1846 /* No worries, td_io_getevents fixes min and max if they are
1847 * set incorrectly */
1848 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1850 td_verror(td, -ret, "td_io_getevents");
1855 init_icd(td, &icd, ret);
1856 ios_completed(td, &icd);
1858 td_verror(td, icd.error, "io_u_queued_complete");
1862 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1863 td->bytes_done[ddir] += icd.bytes_done[ddir];
1869 * Call when io_u is really queued, to update the submission latency.
1871 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1873 if (!td->o.disable_slat) {
1874 unsigned long slat_time;
1876 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1881 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1887 * See if we should reuse the last seed, if dedupe is enabled
1889 static struct frand_state *get_buf_state(struct thread_data *td)
1895 if (!td->o.dedupe_percentage)
1896 return &td->buf_state;
1897 else if (td->o.dedupe_percentage == 100) {
1898 frand_copy(&td->buf_state_prev, &td->buf_state);
1899 return &td->buf_state;
1902 frand_max = rand_max(&td->dedupe_state);
1903 r = __rand(&td->dedupe_state);
1904 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
1906 if (v <= td->o.dedupe_percentage)
1907 return &td->buf_state_prev;
1909 return &td->buf_state;
1912 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1914 if (td->o.dedupe_percentage == 100)
1915 frand_copy(rs, &td->buf_state_prev);
1916 else if (rs == &td->buf_state)
1917 frand_copy(&td->buf_state_prev, rs);
1920 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1921 unsigned int max_bs)
1923 struct thread_options *o = &td->o;
1925 if (o->compress_percentage || o->dedupe_percentage) {
1926 unsigned int perc = td->o.compress_percentage;
1927 struct frand_state *rs;
1928 unsigned int left = max_bs;
1929 unsigned int this_write;
1932 rs = get_buf_state(td);
1934 min_write = min(min_write, left);
1937 this_write = min_not_zero(min_write,
1938 td->o.compress_chunk);
1940 fill_random_buf_percentage(rs, buf, perc,
1941 this_write, this_write,
1943 o->buffer_pattern_bytes);
1945 fill_random_buf(rs, buf, min_write);
1946 this_write = min_write;
1951 save_buf_state(td, rs);
1953 } else if (o->buffer_pattern_bytes)
1954 fill_buffer_pattern(td, buf, max_bs);
1955 else if (o->zero_buffers)
1956 memset(buf, 0, max_bs);
1958 fill_random_buf(get_buf_state(td), buf, max_bs);
1962 * "randomly" fill the buffer contents
1964 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1965 unsigned int min_write, unsigned int max_bs)
1967 io_u->buf_filled_len = 0;
1968 fill_io_buffer(td, io_u->buf, min_write, max_bs);